lustre/llite/rw.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/llite/rw.c
  33  *
  34  * Lustre Lite I/O page cache routines shared by different kernel revs
  35  */
  36
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/string.h>
  40 #include <linux/stat.h>
  41 #include <linux/errno.h>
  42 #include <linux/unistd.h>
  43 #include <linux/writeback.h>
  44 #include <asm/uaccess.h>
  45
  46 #include <linux/fs.h>
  47 #include <linux/file.h>
  48 #include <linux/stat.h>
  49 #include <asm/uaccess.h>
  50 #include <linux/mm.h>
  51 #include <linux/pagemap.h>
  52 /* current_is_kswapd() */
  53 #include <linux/swap.h>
  54 #include <linux/task_io_accounting_ops.h>
  55
  56 #define DEBUG_SUBSYSTEM S_LLITE
  57
  58 #include <obd_cksum.h>
  59 #include "llite_internal.h"
  60 #include <lustre_compat.h>
  61
  62 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
  63
  64 /**
  65  * Get readahead pages from the filesystem readahead pool of the client for a
  66  * thread.
  67  *
  68  * /param sbi superblock for filesystem readahead state ll_ra_info
  69  * /param ria per-thread readahead state
  70  * /param pages number of pages requested for readahead for the thread.
  71  *
  72  * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
  73  * It should work well if the ra_max_pages is much greater than the single
  74  * file's read-ahead window, and not too many threads contending for
  75  * these readahead pages.
  76  *
  77  * TODO: There may be a 'global sync problem' if many threads are trying
  78  * to get an ra budget that is larger than the remaining readahead pages
  79  * and reach here at exactly the same time. They will compute /a ret to
  80  * consume the remaining pages, but will fail at atomic_add_return() and
  81  * get a zero ra window, although there is still ra space remaining. - Jay */
  82
  83 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
  84                                      struct ra_io_arg *ria,
  85                                      unsigned long pages,
  86                                      unsigned long pages_min)
  87 {
  88         struct ll_ra_info *ra = &sbi->ll_ra_info;
  89         long ret;
  90         ENTRY;
  91
  92         /* If read-ahead pages left are less than 1M, do not do read-ahead,
  93          * otherwise it will form small read RPC(< 1M), which hurt server
  94          * performance a lot. */
  95         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
  96                   pages);
  97         if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
  98                 GOTO(out, ret = 0);
  99
 100         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
 101                 atomic_sub(ret, &ra->ra_cur_pages);
 102                 ret = 0;
 103         }
 104
 105 out:
 106         if (ret < pages_min) {
 107                 /* override ra limit for maximum performance */
 108                 atomic_add(pages_min - ret, &ra->ra_cur_pages);
 109                 ret = pages_min;
 110         }
 111         RETURN(ret);
 112 }
 113
 114 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
 115 {
 116         struct ll_ra_info *ra = &sbi->ll_ra_info;
 117         atomic_sub(pages, &ra->ra_cur_pages);
 118 }
 119
 120 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
 121 {
 122         LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
 123         lprocfs_counter_incr(sbi->ll_ra_stats, which);
 124 }
 125
 126 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
 127 {
 128         return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
 129                 sbi->ll_ra_info.ra_max_pages > 0;
 130 }
 131
 132 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
 133 {
 134         struct ll_sb_info *sbi = ll_i2sbi(inode);
 135         ll_ra_stats_inc_sbi(sbi, which);
 136 }
 137
 138 #define RAS_CDEBUG(ras) \
 139         CDEBUG(D_READA,                                                      \
 140                "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu "     \
 141                "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n",             \
 142                ras->ras_last_read_end_bytes, ras->ras_consecutive_requests,  \
 143                ras->ras_consecutive_bytes, ras->ras_window_start_idx,        \
 144                ras->ras_window_pages, ras->ras_next_readahead_idx,           \
 145                ras->ras_rpc_pages, ras->ras_requests,                        \
 146                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
 147                ras->ras_stride_bytes, ras->ras_stride_length,                \
 148                ras->ras_async_last_readpage_idx)
 149
 150 static bool pos_in_window(loff_t pos, loff_t point,
 151                           unsigned long before, unsigned long after)
 152 {
 153         loff_t start = point - before;
 154         loff_t end = point + after;
 155
 156         if (start > point)
 157                 start = 0;
 158         if (end < point)
 159                 end = ~0;
 160
 161         return start <= pos && pos <= end;
 162 }
 163
 164 enum ll_ra_page_hint {
 165         MAYNEED = 0, /* this page possibly accessed soon */
 166         WILLNEED /* this page is gurateed to be needed */
 167 };
 168
 169 /**
 170  * Initiates read-ahead of a page with given index.
 171  *
 172  * \retval +ve: page was already uptodate so it will be skipped
 173  *              from being added;
 174  * \retval -ve: page wasn't added to \a queue for error;
 175  * \retval   0: page was added into \a queue for read ahead.
 176  */
 177 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 178                               struct cl_page_list *queue, pgoff_t index,
 179                               enum ll_ra_page_hint hint)
 180 {
 181         struct cl_object *clob  = io->ci_obj;
 182         struct inode     *inode = vvp_object_inode(clob);
 183         struct page      *vmpage = NULL;
 184         struct cl_page   *page;
 185         struct vvp_page  *vpg;
 186         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 187         int               rc    = 0;
 188         const char       *msg   = NULL;
 189
 190         ENTRY;
 191
 192         switch (hint) {
 193         case MAYNEED:
 194                 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
 195                 if (vmpage == NULL) {
 196                         which = RA_STAT_FAILED_GRAB_PAGE;
 197                         msg   = "g_c_p_n failed";
 198                         GOTO(out, rc = -EBUSY);
 199                 }
 200                 break;
 201         case WILLNEED:
 202                 vmpage = find_or_create_page(inode->i_mapping, index,
 203                                              GFP_NOFS);
 204                 if (vmpage == NULL)
 205                         GOTO(out, rc = -ENOMEM);
 206                 break;
 207         default:
 208                 /* should not come here */
 209                 GOTO(out, rc = -EINVAL);
 210         }
 211
 212         /* Check if vmpage was truncated or reclaimed */
 213         if (vmpage->mapping != inode->i_mapping) {
 214                 which = RA_STAT_WRONG_GRAB_PAGE;
 215                 msg   = "g_c_p_n returned invalid page";
 216                 GOTO(out, rc = -EBUSY);
 217         }
 218
 219         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 220         if (IS_ERR(page)) {
 221                 which = RA_STAT_FAILED_GRAB_PAGE;
 222                 msg   = "cl_page_find failed";
 223                 GOTO(out, rc = PTR_ERR(page));
 224         }
 225
 226         lu_ref_add(&page->cp_reference, "ra", current);
 227         cl_page_assume(env, io, page);
 228         vpg = cl2vvp_page(cl_object_page_slice(clob, page));
 229         if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
 230                 vpg->vpg_defer_uptodate = 1;
 231                 vpg->vpg_ra_used = 0;
 232                 cl_page_list_add(queue, page);
 233         } else {
 234                 /* skip completed pages */
 235                 cl_page_unassume(env, io, page);
 236                 /* This page is already uptodate, returning a positive number
 237                  * to tell the callers about this */
 238                 rc = 1;
 239         }
 240
 241         lu_ref_del(&page->cp_reference, "ra", current);
 242         cl_page_put(env, page);
 243
 244 out:
 245         if (vmpage != NULL) {
 246                 if (rc != 0)
 247                         unlock_page(vmpage);
 248                 put_page(vmpage);
 249         }
 250         if (msg != NULL && hint == MAYNEED) {
 251                 ll_ra_stats_inc(inode, which);
 252                 CDEBUG(D_READA, "%s\n", msg);
 253
 254         }
 255
 256         RETURN(rc);
 257 }
 258
 259 #define RIA_DEBUG(ria)                                                  \
 260         CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n",      \
 261                ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff,    \
 262                ria->ria_length, ria->ria_bytes)
 263
 264 static inline int stride_io_mode(struct ll_readahead_state *ras)
 265 {
 266         return ras->ras_consecutive_stride_requests > 1;
 267 }
 268
 269 /* The function calculates how many bytes will be read in
 270  * [off, off + length], in such stride IO area,
 271  * stride_offset = st_off, stride_lengh = st_len,
 272  * stride_bytes = st_bytes
 273  *
 274  *   |------------------|*****|------------------|*****|------------|*****|....
 275  * st_off
 276  *   |--- st_bytes     ---|
 277  *   |-----     st_len   -----|
 278  *
 279  *              How many bytes it should read in such pattern
 280  *              |-------------------------------------------------------------|
 281  *              off
 282  *              |<------                  length                      ------->|
 283  *
 284  *          =   |<----->|  +  |-------------------------------------| +   |---|
 285  *             start_left                 st_bytes * i                 end_left
 286  */
 287 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
 288                                 loff_t off, loff_t length)
 289 {
 290         u64 start = off > st_off ? off - st_off : 0;
 291         u64 end = off + length > st_off ? off + length - st_off : 0;
 292         u64 start_left;
 293         u64 end_left;
 294         u64 bytes_count;
 295
 296         if (st_len == 0 || length == 0 || end == 0)
 297                 return length;
 298
 299         start = div64_u64_rem(start, st_len, &start_left);
 300         if (start_left < st_bytes)
 301                 start_left = st_bytes - start_left;
 302         else
 303                 start_left = 0;
 304
 305         end = div64_u64_rem(end, st_len, &end_left);
 306         if (end_left > st_bytes)
 307                 end_left = st_bytes;
 308
 309         CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
 310                start, end, start_left, end_left);
 311
 312         if (start == end)
 313                 bytes_count = end_left - (st_bytes - start_left);
 314         else
 315                 bytes_count = start_left +
 316                         st_bytes * (end - start - 1) + end_left;
 317
 318         CDEBUG(D_READA,
 319                "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
 320                st_off, st_len, st_bytes, off, length, bytes_count);
 321
 322         return bytes_count;
 323 }
 324
 325 static unsigned long ria_page_count(struct ra_io_arg *ria)
 326 {
 327         loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
 328                 (loff_t)(ria->ria_end_idx -
 329                          ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
 330         loff_t bytes_count;
 331
 332         if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
 333             (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
 334              ria->ria_stoff & ~PAGE_MASK)) {
 335                 /* Over-estimate un-aligned page stride read */
 336                 unsigned long pg_count = ((ria->ria_bytes +
 337                                            PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
 338                 pg_count *= length_bytes / ria->ria_length + 1;
 339
 340                 return pg_count;
 341         }
 342         bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
 343                                         ria->ria_bytes,
 344                                         (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
 345                                         length_bytes);
 346         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 347 }
 348
 349 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
 350 {
 351         unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
 352
 353         if (opt_size == 0)
 354                 opt_size = 1;
 355         return index - (index % opt_size);
 356 }
 357
 358 /* Check whether the index is in the defined ra-window */
 359 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
 360 {
 361         loff_t pos = (loff_t)idx << PAGE_SHIFT;
 362
 363         /* If ria_length == ria_bytes, it means non-stride I/O mode,
 364          * idx should always inside read-ahead window in this case
 365          * For stride I/O mode, just check whether the idx is inside
 366          * the ria_bytes.
 367          */
 368         if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
 369                 return true;
 370
 371         if (pos >= ria->ria_stoff) {
 372                 u64 offset;
 373
 374                 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
 375
 376                 if (offset < ria->ria_bytes ||
 377                     (ria->ria_length - offset) < PAGE_SIZE)
 378                         return true;
 379         } else if (pos + PAGE_SIZE > ria->ria_stoff) {
 380                 return true;
 381         }
 382
 383         return false;
 384 }
 385
 386 static unsigned long
 387 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
 388                     struct cl_page_list *queue, struct ll_readahead_state *ras,
 389                     struct ra_io_arg *ria, pgoff_t *ra_end)
 390 {
 391         struct cl_read_ahead ra = { 0 };
 392         /* busy page count is per stride */
 393         int rc = 0, count = 0, busy_page_count = 0;
 394         pgoff_t page_idx;
 395
 396         LASSERT(ria != NULL);
 397         RIA_DEBUG(ria);
 398
 399         for (page_idx = ria->ria_start_idx;
 400              page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
 401              page_idx++) {
 402                 if (ras_inside_ra_window(page_idx, ria)) {
 403                         if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
 404                                 pgoff_t end_idx;
 405
 406                                 /*
 407                                  * Do not shrink ria_end_idx at any case until
 408                                  * the minimum end of current read is covered.
 409                                  *
 410                                  * Do not extend read lock accross stripe if
 411                                  * lock contention detected.
 412                                  */
 413                                 if (ra.cra_contention &&
 414                                     page_idx > ria->ria_end_idx_min) {
 415                                         ria->ria_end_idx = *ra_end;
 416                                         break;
 417                                 }
 418
 419                                 cl_read_ahead_release(env, &ra);
 420
 421                                 rc = cl_io_read_ahead(env, io, page_idx, &ra);
 422                                 if (rc < 0)
 423                                         break;
 424
 425                                  /*
 426                                   * Only shrink ria_end_idx if the matched
 427                                   * LDLM lock doesn't cover more.
 428                                   */
 429                                 if (page_idx > ra.cra_end_idx) {
 430                                         ria->ria_end_idx = ra.cra_end_idx;
 431                                         break;
 432                                 }
 433
 434                                 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
 435                                        page_idx, ra.cra_end_idx,
 436                                        ra.cra_rpc_pages);
 437                                 LASSERTF(ra.cra_end_idx >= page_idx,
 438                                          "object: %p, indcies %lu / %lu\n",
 439                                          io->ci_obj, ra.cra_end_idx, page_idx);
 440                                 /* update read ahead RPC size.
 441                                  * NB: it's racy but doesn't matter */
 442                                 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
 443                                     ra.cra_rpc_pages > 0)
 444                                         ras->ras_rpc_pages = ra.cra_rpc_pages;
 445                                 /* trim it to align with optimal RPC size */
 446                                 end_idx = ras_align(ras, ria->ria_end_idx + 1);
 447                                 if (end_idx > 0 && !ria->ria_eof)
 448                                         ria->ria_end_idx = end_idx - 1;
 449                                 if (ria->ria_end_idx < ria->ria_end_idx_min)
 450                                         ria->ria_end_idx = ria->ria_end_idx_min;
 451                         }
 452                         if (page_idx > ria->ria_end_idx)
 453                                 break;
 454
 455                         /* If the page is inside the read-ahead window */
 456                         rc = ll_read_ahead_page(env, io, queue, page_idx,
 457                                                 MAYNEED);
 458                         if (rc < 0 && rc != -EBUSY)
 459                                 break;
 460                         if (rc == -EBUSY) {
 461                                 busy_page_count++;
 462                                 CDEBUG(D_READA,
 463                                        "skip busy page: %lu\n", page_idx);
 464                                 /* For page unaligned readahead the first
 465                                  * last pages of each region can be read by
 466                                  * another reader on the same node, and so
 467                                  * may be busy. So only stop for > 2 busy
 468                                  * pages. */
 469                                 if (busy_page_count > 2)
 470                                         break;
 471                         }
 472
 473                         *ra_end = page_idx;
 474                         /* Only subtract from reserve & count the page if we
 475                          * really did readahead on that page. */
 476                         if (rc == 0) {
 477                                 ria->ria_reserved--;
 478                                 count++;
 479                         }
 480                 } else if (stride_io_mode(ras)) {
 481                         /* If it is not in the read-ahead window, and it is
 482                          * read-ahead mode, then check whether it should skip
 483                          * the stride gap.
 484                          */
 485                         loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
 486                         u64 offset;
 487
 488                         div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
 489                                       &offset);
 490                         if (offset >= ria->ria_bytes) {
 491                                 pos += (ria->ria_length - offset);
 492                                 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
 493                                         page_idx = (pos >> PAGE_SHIFT) - 1;
 494                                 busy_page_count = 0;
 495                                 CDEBUG(D_READA,
 496                                        "Stride: jump %llu pages to %lu\n",
 497                                        ria->ria_length - offset, page_idx);
 498                                 continue;
 499                         }
 500                 }
 501         }
 502
 503         cl_read_ahead_release(env, &ra);
 504
 505         return count;
 506 }
 507
 508 static void ll_readahead_work_free(struct ll_readahead_work *work)
 509 {
 510         fput(work->lrw_file);
 511         OBD_FREE_PTR(work);
 512 }
 513
 514 static void ll_readahead_handle_work(struct work_struct *wq);
 515 static void ll_readahead_work_add(struct inode *inode,
 516                                   struct ll_readahead_work *work)
 517 {
 518         INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
 519         queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
 520                    &work->lrw_readahead_work);
 521 }
 522
 523 static int ll_readahead_file_kms(const struct lu_env *env,
 524                                 struct cl_io *io, __u64 *kms)
 525 {
 526         struct cl_object *clob;
 527         struct inode *inode;
 528         struct cl_attr *attr = vvp_env_thread_attr(env);
 529         int ret;
 530
 531         clob = io->ci_obj;
 532         inode = vvp_object_inode(clob);
 533
 534         cl_object_attr_lock(clob);
 535         ret = cl_object_attr_get(env, clob, attr);
 536         cl_object_attr_unlock(clob);
 537
 538         if (ret != 0)
 539                 RETURN(ret);
 540
 541         *kms = attr->cat_kms;
 542         return 0;
 543 }
 544
 545 static void ll_readahead_handle_work(struct work_struct *wq)
 546 {
 547         struct ll_readahead_work *work;
 548         struct lu_env *env;
 549         __u16 refcheck;
 550         struct ra_io_arg *ria;
 551         struct inode *inode;
 552         struct ll_file_data *fd;
 553         struct ll_readahead_state *ras;
 554         struct cl_io *io;
 555         struct cl_2queue *queue;
 556         pgoff_t ra_end_idx = 0;
 557         unsigned long pages, pages_min = 0;
 558         struct file *file;
 559         __u64 kms;
 560         int rc;
 561         pgoff_t eof_index;
 562         struct ll_sb_info *sbi;
 563
 564         work = container_of(wq, struct ll_readahead_work,
 565                             lrw_readahead_work);
 566         fd = work->lrw_file->private_data;
 567         ras = &fd->fd_ras;
 568         file = work->lrw_file;
 569         inode = file_inode(file);
 570         sbi = ll_i2sbi(inode);
 571
 572         env = cl_env_alloc(&refcheck, LCT_NOREF);
 573         if (IS_ERR(env))
 574                 GOTO(out_free_work, rc = PTR_ERR(env));
 575
 576         io = vvp_env_thread_io(env);
 577         ll_io_init(io, file, CIT_READ, NULL);
 578
 579         rc = ll_readahead_file_kms(env, io, &kms);
 580         if (rc != 0)
 581                 GOTO(out_put_env, rc);
 582
 583         if (kms == 0) {
 584                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 585                 GOTO(out_put_env, rc = 0);
 586         }
 587
 588         ria = &ll_env_info(env)->lti_ria;
 589         memset(ria, 0, sizeof(*ria));
 590
 591         ria->ria_start_idx = work->lrw_start_idx;
 592         /* Truncate RA window to end of file */
 593         eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
 594         if (eof_index <= work->lrw_end_idx) {
 595                 work->lrw_end_idx = eof_index;
 596                 ria->ria_eof = true;
 597         }
 598         if (work->lrw_end_idx <= work->lrw_start_idx)
 599                 GOTO(out_put_env, rc = 0);
 600
 601         ria->ria_end_idx = work->lrw_end_idx;
 602         pages = ria->ria_end_idx - ria->ria_start_idx + 1;
 603         ria->ria_reserved = ll_ra_count_get(sbi, ria,
 604                                             ria_page_count(ria), pages_min);
 605
 606         CDEBUG(D_READA,
 607                "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 608                ria->ria_reserved, pages, pages_min,
 609                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 610                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 611
 612         if (ria->ria_reserved < pages) {
 613                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 614                 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
 615                         ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 616                         GOTO(out_put_env, rc = 0);
 617                 }
 618         }
 619
 620         rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
 621         if (rc)
 622                 GOTO(out_put_env, rc);
 623
 624         /* overwrite jobid inited in vvp_io_init() */
 625         if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 626                     sizeof(work->lrw_jobid)))
 627                 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 628                        sizeof(work->lrw_jobid));
 629
 630         vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
 631         vvp_env_io(env)->vui_fd = fd;
 632         io->ci_state = CIS_LOCKED;
 633         io->ci_async_readahead = true;
 634         rc = cl_io_start(env, io);
 635         if (rc)
 636                 GOTO(out_io_fini, rc);
 637
 638         queue = &io->ci_queue;
 639         cl_2queue_init(queue);
 640
 641         rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
 642                                  &ra_end_idx);
 643         if (ria->ria_reserved != 0)
 644                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 645         if (queue->c2_qin.pl_nr > 0) {
 646                 int count = queue->c2_qin.pl_nr;
 647
 648                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
 649                 if (rc == 0)
 650                         task_io_account_read(PAGE_SIZE * count);
 651         }
 652         if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 653                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 654
 655         if (ra_end_idx != ria->ria_end_idx)
 656                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 657
 658         /* TODO: discard all pages until page reinit route is implemented */
 659         cl_page_list_discard(env, io, &queue->c2_qin);
 660
 661         /* Unlock unsent read pages in case of error. */
 662         cl_page_list_disown(env, io, &queue->c2_qin);
 663
 664         cl_2queue_fini(env, queue);
 665 out_io_fini:
 666         cl_io_end(env, io);
 667         cl_io_fini(env, io);
 668 out_put_env:
 669         cl_env_put(env, &refcheck);
 670 out_free_work:
 671         if (ra_end_idx > 0)
 672                 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
 673         atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
 674         ll_readahead_work_free(work);
 675 }
 676
 677 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
 678                         struct cl_page_list *queue,
 679                         struct ll_readahead_state *ras, bool hit,
 680                         struct file *file)
 681 {
 682         struct vvp_io *vio = vvp_env_io(env);
 683         struct ll_thread_info *lti = ll_env_info(env);
 684         unsigned long pages, pages_min = 0;
 685         pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
 686         struct inode *inode;
 687         struct ra_io_arg *ria = &lti->lti_ria;
 688         struct cl_object *clob;
 689         int ret = 0;
 690         __u64 kms;
 691         ENTRY;
 692
 693         clob = io->ci_obj;
 694         inode = vvp_object_inode(clob);
 695
 696         memset(ria, 0, sizeof(*ria));
 697         ret = ll_readahead_file_kms(env, io, &kms);
 698         if (ret != 0)
 699                 RETURN(ret);
 700
 701         if (kms == 0) {
 702                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 703                 RETURN(0);
 704         }
 705
 706         spin_lock(&ras->ras_lock);
 707
 708         /**
 709          * Note: other thread might rollback the ras_next_readahead_idx,
 710          * if it can not get the full size of prepared pages, see the
 711          * end of this function. For stride read ahead, it needs to
 712          * make sure the offset is no less than ras_stride_offset,
 713          * so that stride read ahead can work correctly.
 714          */
 715         if (stride_io_mode(ras))
 716                 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
 717                                   ras->ras_stride_offset >> PAGE_SHIFT);
 718         else
 719                 start_idx = ras->ras_next_readahead_idx;
 720
 721         if (ras->ras_window_pages > 0)
 722                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
 723
 724         /* Enlarge the RA window to encompass the full read */
 725         if (vio->vui_ra_valid &&
 726             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
 727                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 728
 729         if (end_idx != 0) {
 730                 pgoff_t eof_index;
 731
 732                 /* Truncate RA window to end of file */
 733                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
 734                 if (eof_index <= end_idx) {
 735                         end_idx = eof_index;
 736                         ria->ria_eof = true;
 737                 }
 738         }
 739         ria->ria_start_idx = start_idx;
 740         ria->ria_end_idx = end_idx;
 741         /* If stride I/O mode is detected, get stride window*/
 742         if (stride_io_mode(ras)) {
 743                 ria->ria_stoff = ras->ras_stride_offset;
 744                 ria->ria_length = ras->ras_stride_length;
 745                 ria->ria_bytes = ras->ras_stride_bytes;
 746         }
 747         spin_unlock(&ras->ras_lock);
 748
 749         if (end_idx == 0) {
 750                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 751                 RETURN(0);
 752         }
 753         pages = ria_page_count(ria);
 754         if (pages == 0) {
 755                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 756                 RETURN(0);
 757         }
 758
 759         RAS_CDEBUG(ras);
 760         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
 761                PFID(lu_object_fid(&clob->co_lu)),
 762                ria->ria_start_idx, ria->ria_end_idx,
 763                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
 764                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
 765                hit);
 766
 767         /* at least to extend the readahead window to cover current read */
 768         if (!hit && vio->vui_ra_valid &&
 769             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
 770                 ria->ria_end_idx_min =
 771                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 772                 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
 773                                 ria->ria_start_idx;
 774         }
 775
 776         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
 777                                             pages_min);
 778         if (ria->ria_reserved < pages)
 779                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 780
 781         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 782                ria->ria_reserved, pages, pages_min,
 783                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 784                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 785
 786         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
 787
 788         if (ria->ria_reserved != 0)
 789                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 790
 791         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 792                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 793
 794         CDEBUG(D_READA,
 795                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
 796                ra_end_idx, end_idx, ria->ria_end_idx, ret);
 797
 798         if (ra_end_idx != end_idx)
 799                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 800         if (ra_end_idx > 0) {
 801                 /* update the ras so that the next read-ahead tries from
 802                  * where we left off. */
 803                 spin_lock(&ras->ras_lock);
 804                 ras->ras_next_readahead_idx = ra_end_idx + 1;
 805                 spin_unlock(&ras->ras_lock);
 806                 RAS_CDEBUG(ras);
 807         }
 808
 809         RETURN(ret);
 810 }
 811
 812 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
 813                         struct cl_page_list *queue,
 814                         pgoff_t start, pgoff_t end)
 815 {
 816         int ret = 0;
 817         __u64 kms;
 818         pgoff_t page_idx;
 819         int count = 0;
 820
 821         ENTRY;
 822
 823         ret = ll_readahead_file_kms(env, io, &kms);
 824         if (ret != 0)
 825                 RETURN(ret);
 826
 827         if (kms == 0)
 828                 RETURN(0);
 829
 830         if (end != 0) {
 831                 unsigned long end_index;
 832
 833                 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
 834                 if (end_index <= end)
 835                         end = end_index;
 836         }
 837
 838         for (page_idx = start; page_idx <= end; page_idx++) {
 839                 ret= ll_read_ahead_page(env, io, queue, page_idx,
 840                                         WILLNEED);
 841                 if (ret < 0)
 842                         break;
 843                 else if (ret == 0) /* ret 1 is already uptodate */
 844                         count++;
 845         }
 846
 847         RETURN(count > 0 ? count : ret);
 848 }
 849
 850 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
 851 {
 852         ras->ras_window_start_idx = ras_align(ras, index);
 853 }
 854
 855 /* called with the ras_lock held or from places where it doesn't matter */
 856 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
 857 {
 858         ras->ras_consecutive_requests = 0;
 859         ras->ras_consecutive_bytes = 0;
 860         ras->ras_window_pages = 0;
 861         ras_set_start(ras, index);
 862         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
 863
 864         RAS_CDEBUG(ras);
 865 }
 866
 867 /* called with the ras_lock held or from places where it doesn't matter */
 868 static void ras_stride_reset(struct ll_readahead_state *ras)
 869 {
 870         ras->ras_consecutive_stride_requests = 0;
 871         ras->ras_stride_length = 0;
 872         ras->ras_stride_bytes = 0;
 873         RAS_CDEBUG(ras);
 874 }
 875
 876 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 877 {
 878         spin_lock_init(&ras->ras_lock);
 879         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
 880         ras_reset(ras, 0);
 881         ras->ras_last_read_end_bytes = 0;
 882         ras->ras_requests = 0;
 883 }
 884
 885 /*
 886  * Check whether the read request is in the stride window.
 887  * If it is in the stride window, return true, otherwise return false.
 888  */
 889 static bool read_in_stride_window(struct ll_readahead_state *ras,
 890                                   loff_t pos, loff_t count)
 891 {
 892         loff_t stride_gap;
 893
 894         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
 895             ras->ras_stride_bytes == ras->ras_stride_length)
 896                 return false;
 897
 898         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 899
 900         /* If it is contiguous read */
 901         if (stride_gap == 0)
 902                 return ras->ras_consecutive_bytes + count <=
 903                         ras->ras_stride_bytes;
 904
 905         /* Otherwise check the stride by itself */
 906         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
 907                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
 908                 count <= ras->ras_stride_bytes;
 909 }
 910
 911 static void ras_init_stride_detector(struct ll_readahead_state *ras,
 912                                      loff_t pos, loff_t count)
 913 {
 914         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 915
 916         LASSERT(ras->ras_consecutive_stride_requests == 0);
 917
 918         if (pos <= ras->ras_last_read_end_bytes) {
 919                 /*Reset stride window for forward read*/
 920                 ras_stride_reset(ras);
 921                 return;
 922         }
 923
 924         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
 925         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
 926         ras->ras_consecutive_stride_requests++;
 927         ras->ras_stride_offset = pos;
 928
 929         RAS_CDEBUG(ras);
 930 }
 931
 932 static unsigned long
 933 stride_page_count(struct ll_readahead_state *ras, loff_t len)
 934 {
 935         loff_t bytes_count =
 936                 stride_byte_count(ras->ras_stride_offset,
 937                                   ras->ras_stride_length, ras->ras_stride_bytes,
 938                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
 939
 940         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 941 }
 942
 943 /* Stride Read-ahead window will be increased inc_len according to
 944  * stride I/O pattern */
 945 static void ras_stride_increase_window(struct ll_readahead_state *ras,
 946                                        struct ll_ra_info *ra, loff_t inc_bytes)
 947 {
 948         loff_t window_bytes, stride_bytes;
 949         u64 left_bytes;
 950         u64 step;
 951         loff_t end;
 952
 953         /* temporarily store in page units to reduce LASSERT() cost below */
 954         end = ras->ras_window_start_idx + ras->ras_window_pages;
 955
 956         LASSERT(ras->ras_stride_length > 0);
 957         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
 958                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
 959                  ras->ras_window_start_idx, ras->ras_window_pages,
 960                  ras->ras_stride_offset);
 961
 962         end <<= PAGE_SHIFT;
 963         if (end <= ras->ras_stride_offset)
 964                 stride_bytes = 0;
 965         else
 966                 stride_bytes = end - ras->ras_stride_offset;
 967
 968         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
 969         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
 970         if (left_bytes < ras->ras_stride_bytes) {
 971                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
 972                         window_bytes += inc_bytes;
 973                         goto out;
 974                 } else {
 975                         window_bytes += (ras->ras_stride_bytes - left_bytes);
 976                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
 977                 }
 978         } else {
 979                 window_bytes += (ras->ras_stride_length - left_bytes);
 980         }
 981
 982         LASSERT(ras->ras_stride_bytes != 0);
 983
 984         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
 985
 986         window_bytes += step * ras->ras_stride_length + left_bytes;
 987         LASSERT(window_bytes > 0);
 988
 989 out:
 990         if (stride_page_count(ras, window_bytes) <=
 991             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
 992                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
 993
 994         LASSERT(ras->ras_window_pages > 0);
 995
 996         RAS_CDEBUG(ras);
 997 }
 998
 999 static void ras_increase_window(struct inode *inode,
1000                                 struct ll_readahead_state *ras,
1001                                 struct ll_ra_info *ra)
1002 {
1003         /* The stretch of ra-window should be aligned with max rpc_size
1004          * but current clio architecture does not support retrieve such
1005          * information from lower layer. FIXME later
1006          */
1007         if (stride_io_mode(ras)) {
1008                 ras_stride_increase_window(ras, ra,
1009                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1010         } else {
1011                 pgoff_t window_pages;
1012
1013                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1014                                    ra->ra_max_pages_per_file);
1015                 if (window_pages < ras->ras_rpc_pages)
1016                         ras->ras_window_pages = window_pages;
1017                 else
1018                         ras->ras_window_pages = ras_align(ras, window_pages);
1019         }
1020 }
1021
1022 /**
1023  * Seek within 8 pages are considered as sequential read for now.
1024  */
1025 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1026 {
1027         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1028                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1029 }
1030
1031 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1032                                     struct ll_sb_info *sbi,
1033                                     loff_t pos, size_t count, bool mmap)
1034 {
1035         bool stride_detect = false;
1036         pgoff_t index = pos >> PAGE_SHIFT;
1037
1038         /*
1039          * Reset the read-ahead window in two cases. First when the app seeks
1040          * or reads to some other part of the file. Secondly if we get a
1041          * read-ahead miss that we think we've previously issued. This can
1042          * be a symptom of there being so many read-ahead pages that the VM
1043          * is reclaiming it before we get to it.
1044          */
1045         if (!is_loose_seq_read(ras, pos)) {
1046                 /* Check whether it is in stride I/O mode */
1047                 if (!read_in_stride_window(ras, pos, count)) {
1048                         if (ras->ras_consecutive_stride_requests == 0)
1049                                 ras_init_stride_detector(ras, pos, count);
1050                         else
1051                                 ras_stride_reset(ras);
1052                         ras->ras_consecutive_bytes = 0;
1053                         ras_reset(ras, index);
1054                 } else {
1055                         ras->ras_consecutive_bytes = 0;
1056                         ras->ras_consecutive_requests = 0;
1057                         if (++ras->ras_consecutive_stride_requests > 1)
1058                                 stride_detect = true;
1059                         RAS_CDEBUG(ras);
1060                 }
1061                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1062         } else if (stride_io_mode(ras)) {
1063                 /*
1064                  * If this is contiguous read but in stride I/O mode
1065                  * currently, check whether stride step still is valid,
1066                  * if invalid, it will reset the stride ra window to
1067                  * be zero.
1068                  */
1069                 if (!read_in_stride_window(ras, pos, count)) {
1070                         ras_stride_reset(ras);
1071                         ras->ras_window_pages = 0;
1072                         ras->ras_next_readahead_idx = index;
1073                 }
1074         }
1075
1076         ras->ras_consecutive_bytes += count;
1077         if (mmap) {
1078                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1079
1080                 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1081                         ras->ras_need_increase_window = true;
1082         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1083                 ras->ras_need_increase_window = true;
1084         }
1085
1086         ras->ras_last_read_end_bytes = pos + count - 1;
1087 }
1088
1089 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1090 {
1091         struct ll_file_data *fd = f->private_data;
1092         struct ll_readahead_state *ras = &fd->fd_ras;
1093         struct inode *inode = file_inode(f);
1094         unsigned long index = pos >> PAGE_SHIFT;
1095         struct ll_sb_info *sbi = ll_i2sbi(inode);
1096
1097         spin_lock(&ras->ras_lock);
1098         ras->ras_requests++;
1099         ras->ras_consecutive_requests++;
1100         ras->ras_need_increase_window = false;
1101         ras->ras_no_miss_check = false;
1102         /*
1103          * On the second access to a file smaller than the tunable
1104          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1105          * file up to ra_max_pages_per_file.  This is simply a best effort
1106          * and only occurs once per open file. Normal RA behavior is reverted
1107          * to for subsequent IO.
1108          */
1109         if (ras->ras_requests >= 2) {
1110                 __u64 kms_pages;
1111                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1112
1113                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1114                             PAGE_SHIFT;
1115
1116                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1117                        ra->ra_max_read_ahead_whole_pages,
1118                        ra->ra_max_pages_per_file);
1119
1120                 if (kms_pages &&
1121                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1122                         ras->ras_window_start_idx = 0;
1123                         ras->ras_next_readahead_idx = index + 1;
1124                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1125                                             ra->ra_max_read_ahead_whole_pages);
1126                         ras->ras_no_miss_check = true;
1127                         GOTO(out_unlock, 0);
1128                 }
1129         }
1130         ras_detect_read_pattern(ras, sbi, pos, count, false);
1131 out_unlock:
1132         spin_unlock(&ras->ras_lock);
1133 }
1134
1135 static bool index_in_stride_window(struct ll_readahead_state *ras,
1136                                    pgoff_t index)
1137 {
1138         loff_t pos = (loff_t)index << PAGE_SHIFT;
1139
1140         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1141             ras->ras_stride_bytes == ras->ras_stride_length)
1142                 return false;
1143
1144         if (pos >= ras->ras_stride_offset) {
1145                 u64 offset;
1146
1147                 div64_u64_rem(pos - ras->ras_stride_offset,
1148                               ras->ras_stride_length, &offset);
1149                 if (offset < ras->ras_stride_bytes ||
1150                     ras->ras_stride_length - offset < PAGE_SIZE)
1151                         return true;
1152         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1153                 return true;
1154         }
1155
1156         return false;
1157 }
1158
1159 /*
1160  * ll_ras_enter() is used to detect read pattern according to pos and count.
1161  *
1162  * ras_update() is used to detect cache miss and
1163  * reset window or increase window accordingly
1164  */
1165 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1166                        struct ll_readahead_state *ras, pgoff_t index,
1167                        enum ras_update_flags flags)
1168 {
1169         struct ll_ra_info *ra = &sbi->ll_ra_info;
1170         bool hit = flags & LL_RAS_HIT;
1171
1172         ENTRY;
1173         spin_lock(&ras->ras_lock);
1174
1175         if (!hit)
1176                 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1177                        PFID(ll_inode2fid(inode)), index);
1178         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1179
1180         /*
1181          * The readahead window has been expanded to cover whole
1182          * file size, we don't care whether ra miss happen or not.
1183          * Because we will read whole file to page cache even if
1184          * some pages missed.
1185          */
1186         if (ras->ras_no_miss_check)
1187                 GOTO(out_unlock, 0);
1188
1189         if (flags & LL_RAS_MMAP)
1190                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1191                                         PAGE_SIZE, true);
1192
1193         if (!hit && ras->ras_window_pages &&
1194             index < ras->ras_next_readahead_idx &&
1195             pos_in_window(index, ras->ras_window_start_idx, 0,
1196                           ras->ras_window_pages)) {
1197                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1198                 ras->ras_need_increase_window = false;
1199
1200                 if (index_in_stride_window(ras, index) &&
1201                     stride_io_mode(ras)) {
1202                         /*
1203                          * if (index != ras->ras_last_readpage + 1)
1204                          *      ras->ras_consecutive_pages = 0;
1205                          */
1206                         ras_reset(ras, index);
1207
1208                         /*
1209                          * If stride-RA hit cache miss, the stride
1210                          * detector will not be reset to avoid the
1211                          * overhead of redetecting read-ahead mode,
1212                          * but on the condition that the stride window
1213                          * is still intersect with normal sequential
1214                          * read-ahead window.
1215                          */
1216                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1217                                 ras_stride_reset(ras);
1218                         RAS_CDEBUG(ras);
1219                 } else {
1220                         /*
1221                          * Reset both stride window and normal RA
1222                          * window.
1223                          */
1224                         ras_reset(ras, index);
1225                         /* ras->ras_consecutive_pages++; */
1226                         ras->ras_consecutive_bytes = 0;
1227                         ras_stride_reset(ras);
1228                         GOTO(out_unlock, 0);
1229                 }
1230         }
1231         ras_set_start(ras, index);
1232
1233         if (stride_io_mode(ras)) {
1234                 /* Since stride readahead is sentivite to the offset
1235                  * of read-ahead, so we use original offset here,
1236                  * instead of ras_window_start_idx, which is RPC aligned.
1237                  */
1238                 ras->ras_next_readahead_idx = max(index + 1,
1239                                                   ras->ras_next_readahead_idx);
1240                 ras->ras_window_start_idx =
1241                                 max_t(pgoff_t, ras->ras_window_start_idx,
1242                                       ras->ras_stride_offset >> PAGE_SHIFT);
1243         } else {
1244                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1245                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1246                 if (!hit)
1247                         ras->ras_next_readahead_idx = index + 1;
1248         }
1249
1250         if (ras->ras_need_increase_window) {
1251                 ras_increase_window(inode, ras, ra);
1252                 ras->ras_need_increase_window = false;
1253         }
1254
1255         EXIT;
1256 out_unlock:
1257         spin_unlock(&ras->ras_lock);
1258 }
1259
1260 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1261 {
1262         struct inode           *inode = vmpage->mapping->host;
1263         struct ll_inode_info   *lli   = ll_i2info(inode);
1264         struct lu_env          *env;
1265         struct cl_io           *io;
1266         struct cl_page         *page;
1267         struct cl_object       *clob;
1268         bool redirtied = false;
1269         bool unlocked = false;
1270         int result;
1271         __u16 refcheck;
1272         ENTRY;
1273
1274         LASSERT(PageLocked(vmpage));
1275         LASSERT(!PageWriteback(vmpage));
1276
1277         LASSERT(ll_i2dtexp(inode) != NULL);
1278
1279         env = cl_env_get(&refcheck);
1280         if (IS_ERR(env))
1281                 GOTO(out, result = PTR_ERR(env));
1282
1283         clob  = ll_i2info(inode)->lli_clob;
1284         LASSERT(clob != NULL);
1285
1286         io = vvp_env_thread_io(env);
1287         io->ci_obj = clob;
1288         io->ci_ignore_layout = 1;
1289         result = cl_io_init(env, io, CIT_MISC, clob);
1290         if (result == 0) {
1291                 page = cl_page_find(env, clob, vmpage->index,
1292                                     vmpage, CPT_CACHEABLE);
1293                 if (!IS_ERR(page)) {
1294                         lu_ref_add(&page->cp_reference, "writepage",
1295                                    current);
1296                         cl_page_assume(env, io, page);
1297                         result = cl_page_flush(env, io, page);
1298                         if (result != 0) {
1299                                 /*
1300                                  * Re-dirty page on error so it retries write,
1301                                  * but not in case when IO has actually
1302                                  * occurred and completed with an error.
1303                                  */
1304                                 if (!PageError(vmpage)) {
1305                                         redirty_page_for_writepage(wbc, vmpage);
1306                                         result = 0;
1307                                         redirtied = true;
1308                                 }
1309                         }
1310                         cl_page_disown(env, io, page);
1311                         unlocked = true;
1312                         lu_ref_del(&page->cp_reference,
1313                                    "writepage", current);
1314                         cl_page_put(env, page);
1315                 } else {
1316                         result = PTR_ERR(page);
1317                 }
1318         }
1319         cl_io_fini(env, io);
1320
1321         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1322                 loff_t offset = cl_offset(clob, vmpage->index);
1323
1324                 /* Flush page failed because the extent is being written out.
1325                  * Wait for the write of extent to be finished to avoid
1326                  * breaking kernel which assumes ->writepage should mark
1327                  * PageWriteback or clean the page. */
1328                 result = cl_sync_file_range(inode, offset,
1329                                             offset + PAGE_SIZE - 1,
1330                                             CL_FSYNC_LOCAL, 1);
1331                 if (result > 0) {
1332                         /* actually we may have written more than one page.
1333                          * decreasing this page because the caller will count
1334                          * it. */
1335                         wbc->nr_to_write -= result - 1;
1336                         result = 0;
1337                 }
1338         }
1339
1340         cl_env_put(env, &refcheck);
1341         GOTO(out, result);
1342
1343 out:
1344         if (result < 0) {
1345                 if (!lli->lli_async_rc)
1346                         lli->lli_async_rc = result;
1347                 SetPageError(vmpage);
1348                 if (!unlocked)
1349                         unlock_page(vmpage);
1350         }
1351         return result;
1352 }
1353
1354 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1355 {
1356         struct inode *inode = mapping->host;
1357         loff_t start;
1358         loff_t end;
1359         enum cl_fsync_mode mode;
1360         int range_whole = 0;
1361         int result;
1362         ENTRY;
1363
1364         if (wbc->range_cyclic) {
1365                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1366                 end = OBD_OBJECT_EOF;
1367         } else {
1368                 start = wbc->range_start;
1369                 end = wbc->range_end;
1370                 if (end == LLONG_MAX) {
1371                         end = OBD_OBJECT_EOF;
1372                         range_whole = start == 0;
1373                 }
1374         }
1375
1376         mode = CL_FSYNC_NONE;
1377         if (wbc->sync_mode == WB_SYNC_ALL)
1378                 mode = CL_FSYNC_LOCAL;
1379
1380         if (ll_i2info(inode)->lli_clob == NULL)
1381                 RETURN(0);
1382
1383         /* for directio, it would call writepages() to evict cached pages
1384          * inside the IO context of write, which will cause deadlock at
1385          * layout_conf since it waits for active IOs to complete. */
1386         result = cl_sync_file_range(inode, start, end, mode, 1);
1387         if (result > 0) {
1388                 wbc->nr_to_write -= result;
1389                 result = 0;
1390          }
1391
1392         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1393                 if (end == OBD_OBJECT_EOF)
1394                         mapping->writeback_index = 0;
1395                 else
1396                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1397         }
1398         RETURN(result);
1399 }
1400
1401 struct ll_cl_context *ll_cl_find(struct file *file)
1402 {
1403         struct ll_file_data *fd = file->private_data;
1404         struct ll_cl_context *lcc;
1405         struct ll_cl_context *found = NULL;
1406
1407         read_lock(&fd->fd_lock);
1408         list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1409                 if (lcc->lcc_cookie == current) {
1410                         found = lcc;
1411                         break;
1412                 }
1413         }
1414         read_unlock(&fd->fd_lock);
1415
1416         return found;
1417 }
1418
1419 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1420                enum lcc_type type)
1421 {
1422         struct ll_file_data *fd = file->private_data;
1423         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1424
1425         memset(lcc, 0, sizeof(*lcc));
1426         INIT_LIST_HEAD(&lcc->lcc_list);
1427         lcc->lcc_cookie = current;
1428         lcc->lcc_env = env;
1429         lcc->lcc_io = io;
1430         lcc->lcc_type = type;
1431
1432         write_lock(&fd->fd_lock);
1433         list_add(&lcc->lcc_list, &fd->fd_lccs);
1434         write_unlock(&fd->fd_lock);
1435 }
1436
1437 void ll_cl_remove(struct file *file, const struct lu_env *env)
1438 {
1439         struct ll_file_data *fd = file->private_data;
1440         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1441
1442         write_lock(&fd->fd_lock);
1443         list_del_init(&lcc->lcc_list);
1444         write_unlock(&fd->fd_lock);
1445 }
1446
1447 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1448                            struct cl_page *page, struct file *file)
1449 {
1450         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1451         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1452         struct ll_file_data       *fd     = NULL;
1453         struct ll_readahead_state *ras    = NULL;
1454         struct cl_2queue          *queue  = &io->ci_queue;
1455         struct cl_sync_io         *anchor = NULL;
1456         struct vvp_page           *vpg;
1457         int                        rc = 0, rc2 = 0;
1458         bool                       uptodate;
1459         pgoff_t io_start_index;
1460         pgoff_t io_end_index;
1461         ENTRY;
1462
1463         if (file) {
1464                 fd = file->private_data;
1465                 ras = &fd->fd_ras;
1466         }
1467
1468         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1469         uptodate = vpg->vpg_defer_uptodate;
1470
1471         if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1472                 struct vvp_io *vio = vvp_env_io(env);
1473                 enum ras_update_flags flags = 0;
1474
1475                 if (uptodate)
1476                         flags |= LL_RAS_HIT;
1477                 if (!vio->vui_ra_valid)
1478                         flags |= LL_RAS_MMAP;
1479                 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1480         }
1481
1482         cl_2queue_init(queue);
1483         if (uptodate) {
1484                 vpg->vpg_ra_used = 1;
1485                 cl_page_export(env, page, 1);
1486                 cl_page_disown(env, io, page);
1487         } else {
1488                 anchor = &vvp_env_info(env)->vti_anchor;
1489                 cl_sync_io_init(anchor, 1);
1490                 page->cp_sync_io = anchor;
1491
1492                 cl_2queue_add(queue, page);
1493         }
1494
1495         io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1496         io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1497                                 io->u.ci_rw.crw_count - 1);
1498         if (ll_readahead_enabled(sbi) && ras) {
1499                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1500                                    uptodate, file);
1501                 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1502                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1503         } else if (vvp_index(vpg) == io_start_index &&
1504                    io_end_index - io_start_index > 0) {
1505                 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1506                                    io_end_index);
1507                 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1508                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1509         }
1510
1511         if (queue->c2_qin.pl_nr > 0) {
1512                 int count = queue->c2_qin.pl_nr;
1513                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1514                 if (rc == 0)
1515                         task_io_account_read(PAGE_SIZE * count);
1516         }
1517
1518
1519         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1520                 rc = cl_sync_io_wait(env, anchor, 0);
1521
1522                 cl_page_assume(env, io, page);
1523                 cl_page_list_del(env, &queue->c2_qout, page);
1524
1525                 if (!PageUptodate(cl_page_vmpage(page))) {
1526                         /* Failed to read a mirror, discard this page so that
1527                          * new page can be created with new mirror.
1528                          *
1529                          * TODO: this is not needed after page reinit
1530                          * route is implemented */
1531                         cl_page_discard(env, io, page);
1532                 }
1533                 cl_page_disown(env, io, page);
1534         }
1535
1536         /* TODO: discard all pages until page reinit route is implemented */
1537         cl_page_list_discard(env, io, &queue->c2_qin);
1538
1539         /* Unlock unsent read pages in case of error. */
1540         cl_page_list_disown(env, io, &queue->c2_qin);
1541
1542         cl_2queue_fini(env, queue);
1543
1544         RETURN(rc);
1545 }
1546
1547 /*
1548  * Possible return value:
1549  * 0 no async readahead triggered and fast read could not be used.
1550  * 1 no async readahead, but fast read could be used.
1551  * 2 async readahead triggered and fast read could be used too.
1552  * < 0 on error.
1553  */
1554 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1555 {
1556         struct ll_readahead_work *lrw;
1557         struct inode *inode = file_inode(file);
1558         struct ll_sb_info *sbi = ll_i2sbi(inode);
1559         struct ll_file_data *fd = file->private_data;
1560         struct ll_readahead_state *ras = &fd->fd_ras;
1561         struct ll_ra_info *ra = &sbi->ll_ra_info;
1562         unsigned long throttle;
1563         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1564         pgoff_t end_idx = start_idx + pages - 1;
1565
1566         throttle = min(ra->ra_async_pages_per_file_threshold,
1567                        ra->ra_max_pages_per_file);
1568         /*
1569          * If this is strided i/o or the window is smaller than the
1570          * throttle limit, we do not do async readahead. Otherwise,
1571          * we do async readahead, allowing the user thread to do fast i/o.
1572          */
1573         if (stride_io_mode(ras) || !throttle ||
1574             ras->ras_window_pages < throttle ||
1575             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1576                 return 0;
1577
1578         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1579                 return 0;
1580
1581         if (ras->ras_async_last_readpage_idx == start_idx)
1582                 return 1;
1583
1584         /* ll_readahead_work_free() free it */
1585         OBD_ALLOC_PTR(lrw);
1586         if (lrw) {
1587                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1588                 lrw->lrw_file = get_file(file);
1589                 lrw->lrw_start_idx = start_idx;
1590                 lrw->lrw_end_idx = end_idx;
1591                 spin_lock(&ras->ras_lock);
1592                 ras->ras_next_readahead_idx = end_idx + 1;
1593                 ras->ras_async_last_readpage_idx = start_idx;
1594                 spin_unlock(&ras->ras_lock);
1595                 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1596                        sizeof(lrw->lrw_jobid));
1597                 ll_readahead_work_add(inode, lrw);
1598         } else {
1599                 return -ENOMEM;
1600         }
1601
1602         return 2;
1603 }
1604
1605 /*
1606  * Check if we can issue a readahead RPC, if that is
1607  * the case, we can't do fast IO because we will need
1608  * a cl_io to issue the RPC.
1609  */
1610 static bool ll_use_fast_io(struct file *file,
1611                            struct ll_readahead_state *ras, pgoff_t index)
1612 {
1613         unsigned long fast_read_pages =
1614                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1615         loff_t skip_pages;
1616         loff_t stride_bytes = ras->ras_stride_bytes;
1617
1618         if (stride_io_mode(ras) && stride_bytes) {
1619                 skip_pages = (ras->ras_stride_length +
1620                         ras->ras_stride_bytes - 1) / stride_bytes;
1621                 skip_pages *= fast_read_pages;
1622         } else {
1623                 skip_pages = fast_read_pages;
1624         }
1625
1626         if (ras->ras_window_start_idx + ras->ras_window_pages <
1627             ras->ras_next_readahead_idx + skip_pages ||
1628             kickoff_async_readahead(file, fast_read_pages) > 0)
1629                 return true;
1630
1631         return false;
1632 }
1633
1634 int ll_readpage(struct file *file, struct page *vmpage)
1635 {
1636         struct inode *inode = file_inode(file);
1637         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1638         struct ll_cl_context *lcc;
1639         const struct lu_env  *env = NULL;
1640         struct cl_io   *io = NULL;
1641         struct cl_page *page;
1642         struct ll_sb_info *sbi = ll_i2sbi(inode);
1643         int result;
1644         ENTRY;
1645
1646         lcc = ll_cl_find(file);
1647         if (lcc != NULL) {
1648                 env = lcc->lcc_env;
1649                 io  = lcc->lcc_io;
1650         }
1651
1652         if (io == NULL) { /* fast read */
1653                 struct inode *inode = file_inode(file);
1654                 struct ll_file_data *fd = file->private_data;
1655                 struct ll_readahead_state *ras = &fd->fd_ras;
1656                 struct lu_env  *local_env = NULL;
1657                 struct vvp_page *vpg;
1658
1659                 result = -ENODATA;
1660
1661                 /* TODO: need to verify the layout version to make sure
1662                  * the page is not invalid due to layout change. */
1663                 page = cl_vmpage_page(vmpage, clob);
1664                 if (page == NULL) {
1665                         unlock_page(vmpage);
1666                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1667                         RETURN(result);
1668                 }
1669
1670                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1671                 if (vpg->vpg_defer_uptodate) {
1672                         enum ras_update_flags flags = LL_RAS_HIT;
1673
1674                         if (lcc && lcc->lcc_type == LCC_MMAP)
1675                                 flags |= LL_RAS_MMAP;
1676
1677                         /* For fast read, it updates read ahead state only
1678                          * if the page is hit in cache because non cache page
1679                          * case will be handled by slow read later. */
1680                         ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1681                         /* avoid duplicate ras_update() call */
1682                         vpg->vpg_ra_updated = 1;
1683
1684                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1685                                 result = 0;
1686                 }
1687
1688                 if (!env) {
1689                         local_env = cl_env_percpu_get();
1690                         env = local_env;
1691                 }
1692
1693                 /* export the page and skip io stack */
1694                 if (result == 0) {
1695                         vpg->vpg_ra_used = 1;
1696                         cl_page_export(env, page, 1);
1697                 } else {
1698                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1699                 }
1700                 /* release page refcount before unlocking the page to ensure
1701                  * the object won't be destroyed in the calling path of
1702                  * cl_page_put(). Please see comment in ll_releasepage(). */
1703                 cl_page_put(env, page);
1704                 unlock_page(vmpage);
1705                 if (local_env)
1706                         cl_env_percpu_put(local_env);
1707
1708                 RETURN(result);
1709         }
1710
1711         /**
1712          * Direct read can fall back to buffered read, but DIO is done
1713          * with lockless i/o, and buffered requires LDLM locking, so in
1714          * this case we must restart without lockless.
1715          */
1716         if (file->f_flags & O_DIRECT &&
1717             lcc && lcc->lcc_type == LCC_RW &&
1718             !io->ci_ignore_lockless) {
1719                 unlock_page(vmpage);
1720                 io->ci_ignore_lockless = 1;
1721                 io->ci_need_restart = 1;
1722                 RETURN(-ENOLCK);
1723         }
1724
1725         LASSERT(io->ci_state == CIS_IO_GOING);
1726         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1727         if (!IS_ERR(page)) {
1728                 LASSERT(page->cp_type == CPT_CACHEABLE);
1729                 if (likely(!PageUptodate(vmpage))) {
1730                         cl_page_assume(env, io, page);
1731
1732                         result = ll_io_read_page(env, io, page, file);
1733                 } else {
1734                         /* Page from a non-object file. */
1735                         unlock_page(vmpage);
1736                         result = 0;
1737                 }
1738                 cl_page_put(env, page);
1739         } else {
1740                 unlock_page(vmpage);
1741                 result = PTR_ERR(page);
1742         }
1743         RETURN(result);
1744 }